Source-linked AI summary
Rethinking Multimodal Fake News Detection in the Generative AI Era
Wenbin Shen, Guoxuan Qin, Guangxu Yao, Baodong Wang, Yuanbo Rui, Zhichao Lian
TL;DR
Existing multimodal fake-news datasets rarely characterize generative involvement, while AIGC detection does not establish whether a news event is true. The paper constructs Weibo26 and proposes GAHR, combining global judgment, local correction, and generativity signals. GAHR achieves competitive veracity detection on evaluated datasets and strong generative-content detection on Weibo26.
Problem
Existing multimodal fake-news datasets lack generativity information, while AIGC detection identifies generated content but does not determine news-event truth.
Source: §1 · p. 2Method
The paper constructs Weibo26 with veracity and generativity labels and proposes GAHR, which uses global judgment, local correction, and generativity judgments as an auxiliary signal.
Source: §1 · p. 2Results
GAHR achieves competitive veracity-detection performance on evaluated datasets and strong generative-content detection results on Weibo26.
Source: §6 · p. 18Takeaways & Limitations
Weibo26 provides a setting for jointly evaluating news veracity and generative-content detection when generated content participates in dissemination.
Source: §6 · p. 17The case study reports ICRN and TCFN errors in which coherent image–text pairs grounded in false narratives were classified as real.
Source: §5.8 · pp. 16–17
Abstract
from arXiv · showhide
Generative content is increasingly entering the production and dissemination of news, transforming fake news from manually fabricated or simply manipulated material into complex forms in which native and generated content jointly participate. Existing multimodal fake news detection research primarily focuses on veracity assessment and rarely characterizes how generativity differences affect the reliability of evidence. In contrast, AIGC detection primarily determines whether content is generated or modified by generative models, but it does not by itself establish whether the underlying news event is true. To bridge the separation between these tasks in data and evaluation, we construct Weibo26, a multimodal fake news detection dataset for generative-content scenarios. On this basis, we propose the Generativity-Aware Hierarchical Reasoning (GAHR) framework, which combines global judgment with local correction so that generativity information participates in news-veracity reasoning. Experiments on multiple existing fake news detection benchmarks and Weibo26 show that GAHR achieves competitive veracity-detection performance while effectively identifying generative content.
1. Introduction
Generative content complicates multimodal veracity assessment, while existing veracity and AIGC detection address different questions. This work introduces Weibo26 and GAHR to bring generativity information into veracity evaluation and reasoning.
1. Introduction
- Fake news increasingly combines native and generated content, extending beyond manual fabrication, image–text mismatch, or local manipulation.Generative AI has entered news production and dissemination, with models producing news-style text and visual content.Source: §1 · p. 1
- Existing datasets and methods lack a unified treatment of generativity and veracity: conventional datasets omit generativity labels, while generated-content detectors cannot replace veracity reasoning.Multimodal fake news methods assess veracity, whereas AIGC detection asks whether content was generated; neither task alone covers both dimensions.Source: §1 · pp. 1–2
- Weibo26 contains 15,154 samples across six types, with both veracity and generativity labels for evaluating generative involvement in news-veracity scenarios.It characterizes veracity differences across types of generative involvement.Source: §1 · p. 2
- GAHR forms a global veracity judgment through bidirectional cross-modal interaction, then uses local text, image, and image–text evidence for gated correction.On annotated data, it also learns text- and image-generativity judgments as auxiliary training signals.Source: §1 · p. 2
- Experiments cover five existing multimodal fake-news benchmarks and Weibo26, assessing the dataset and method across veracity detection and AIGC detection.The paper reports systematic experiments from multiple perspectives.Source: §1 · p. 2
2. Related Work
Prior work develops multimodal veracity detection and AIGC detection as distinct lines of research. Their datasets and evaluation objectives do not fully capture generative involvement in news-veracity assessment.
2.1. Multimodal Fake News Detection
- Multimodal fake-news research progressed from single-modality and social-context features to fusion, cross-modal modeling, and newer fine-grained and vision–language methods (Ma et al., 2016, 2018; Ruchansky et al., 2017; Wang et al., 2018; Yu et al., 2026; Liu et al., 2026a).Existing datasets and evaluation protocols primarily target real–fake classification and incompletely cover sample forms and evidence structures involving generative content (Boididou et al., 2015, 2014; Nakamura et al., 2020; Boididou et al., 2018).Source: §2.1 · p. 3
2.2. AIGC Detection
- AIGC detectors use textual statistical cues or visual artifacts and fingerprints, but their evaluations focus on generativity rather than jointly assessing news veracity and generative involvement.Generated-text detector effectiveness is related to the generation model, domain, and post-processing, while detectors based on superficial features may face cross-setting generalization difficulties (Jawahar et al., 2020; Crothers et al., 2023; Sadasivan et al., 2023).Source: §2.2 · p. 3
3. Construction of the Weibo26 Dataset
Weibo26 is built from verified social-media news and manually checked derived samples, with distinct types for native and generative involvement. Its labels and sample distribution support evaluation of veracity alongside generativity.
3.1. Data Collection and Cleaning
- Original real and fake news samples are evidence-screened, cleaned for text and image defects, and retained only after three-reviewer agreement on completeness, classifiability, and veracity.The source criteria require reliable evidence for real news and debunking or fact-checking evidence for fake news; ambiguous samples are excluded or reconstructed and reviewed again.Source: §3.1 · pp. 3–4
3.2. Construction of Generative Samples
- Generated samples comprise image-conditioned fact-consistent or fact-deviating text (ICRN/ICFN) and text-conditioned consistent or misleading images (TCRN/TCFN), followed by manual verification.Selected source samples are not included in the final dataset alongside their derived samples; generated samples are retained only after unanimous reviewer approval.Source: §3.2 · p. 4, §3.1 · p. 4
3.3. Analysis of the Weibo26 Dataset
- 15,154 samples span six types, with 9,473 real and 5,681 fake samples, and 9,571 generative-content versus 5,583 native samples.The labels separately indicate factual consistency and whether a sample contains generative content.Source: §3.3 · p. 5, Figure 2 · p. 5
- Table 1 compares Weibo26 with existing multimodal fake-news datasets by language, publication year, temporal range, and sample scale.The accompanying discussion describes Weibo26 as retaining veracity labels while adding generative-sample construction.Source: Table 1 · p. 5, §3.3 · p. 5
4. Methodology
GAHR combines a global veracity judgment with local evidence-based correction, while using generativity labels as an auxiliary signal when available. Its method projects multimodal features into a shared space and integrates global and local reasoning with generativity detection.
4.1. Overview
- GAHR combines global judgment, local correction, and an auxiliary branch that predicts text- and image-side generative-content status when labels are available.The auxiliary predictions allow generativity information to participate in representation learning for veracity judgment.Source: §4.1 · p. 6
4.2. Unified Encoding Representation
- GAHR projects text, image, and cross-modal representations into a unified dimension because encoder output dimensions are not fully consistent.The unified representation includes global features and local sequences from base and cross-modal encoders.Source: §4.2 · pp. 6–7
4.3. Global Judgment
- Bidirectional cross-modal responses feed global interaction modeling, which combines self-attention outputs with elementwise interactions to produce an overall veracity logit.The global judgment is subsequently adjusted by local correction.Source: §4.3 · pp. 7–8
4.4. Local Correction
- Local correction derives text-side, image-side, and image–text consistency evidence, then adds a learned gated correction to the global judgment as a residual.The first two evidence types characterize within-modality differences between base and cross-modal semantics; the third characterizes local image–text consistency.Source: §4.4 · pp. 9–10
4.5. Generativity Detection
- When generativity labels are available, a branch uses the global representation and three local evidence types to predict text- and image-side generation probabilities.This branch lets generativity information participate in veracity representation learning.Source: §4.5 · p. 10
4.6. Optimization Objective
- Training combines veracity-classification loss with an auxiliary generativity-detection loss, computed for samples carrying generativity annotations.The generativity loss is zero when a batch contains no annotated samples.Source: §4.6 · pp. 10–11
5. Experiments
Experiments evaluate GAHR on public fake-news benchmarks and Weibo26, using dataset-specific splits and baselines for veracity and generativity detection.
5.1. Datasets
- The six-dataset evaluation covers five public datasets and Weibo26, testing conventional veracity detection and, on Weibo26, veracity and generativity recognition amid generative content.Public datasets use uniform cleaning rules to address invalid links, corrupted or undecodable images, and missing text fields.Source: §5.1 · p. 11
- PHEME is split by event, while datasets without official partitions, including Weibo26, use event-level clusters split 3:1:1; official partitions are retained where available.Samples from the same semantic event stay within one partition.Source: §5.1 · p. 11
5.2. Baselines
- Baselines are selected for two tasks: veracity classification and recognition of text- and image-side generativity labels.Veracity baselines span eight multimodal fake-news approaches, including hierarchical attention, cross-modal fusion, and neuro-symbolic reasoning.Source: §5.2 · p. 11
- AIGC baselines cover language-model, text-distribution, and discriminative-pattern methods for text, plus gradient, frequency, artifact-generalization, universal, and neighboring-pixel strategies for images.All methods use the same splits and evaluation metrics; pretrained-feature configurations follow original papers or public implementations.Source: §5.2 · pp. 11–12
5.3. Implementation Details
The implementation uses pretrained text, image, and cross-modal encoders, a shared representation space, and repeated seeded runs.
5.3. Implementation Details
- Experiments project RoBERTa, DINOv2, and CLIP features into 512 dimensions and report mean and standard deviation over five random seeds.Training uses batch size 64, up to 50 epochs, AdamW, a 5×10−5 initial learning rate, weight decay 0.1, cosine annealing with 10% warmup, and early stopping.Source: §5.3 · p. 12
5.4. Veracity Detection Analysis
GAHR is competitive on veracity detection across datasets and performs particularly well on fake-news metrics. On Weibo26, it leads overall despite receiving no generativity supervision in this experiment.
5.4. Veracity Detection Analysis
- GAHR shows strong fake-news Recall or F1-score on PHEME, GossipCop, AMG, and Weibo26, and high fake-news F1-score on Weibo17 and Weibo21.Some baselines have insufficient fake-news recall, particularly on datasets with imbalanced class distributions.Source: §5.4 · p. 12
- On Weibo26, GAHR achieves the best overall Accuracy, fake-news F1-score, and AUC without generativity supervision.Test samples include native news and news involving generative content.Source: §5.4 · p. 12
- GAHR achieves best or near-best Accuracy and AUC on PHEME, AMG, Weibo17, Weibo21, and Weibo26.The comparison uses veracity labels only; the table reports mean ± standard deviation over five runs.Source: §5.4 · p. 12, Table 2 · p. 13
5.5. Ablation Study
Ablations on AMG, Weibo21, and Weibo26 find that the full GAHR model performs best overall. Removing text, images, global judgment, or local correction weakens performance, with both reasoning levels contributing on Weibo26.
5.5. Ablation Study
- The full GAHR model achieves the best overall performance on AMG, Weibo21, and Weibo26.The ablation removes text, images, global judgment, and local correction separately.Source: §5.5 · p. 14
- Removing text causes marked declines across all datasets, images provide supplementary evidence, and removing either reasoning level lowers performance.On Weibo26, the full model exceeds both structural-ablation variants in fake-news Recall and F1-score, indicating complementary roles for global judgment and local correction in samples involving generative content.Source: §5.5 · p. 14
5.6. AIGC Detection Analysis
GAHR performs text- and image-side generativity detection using modality-specific variants, and the reported comparisons favor GAHR in both settings. The analysis also finds that generativity labels improve veracity accuracy across all six sample types, while GAHR processing separates their representations more clearly in a visualization.
5.6. AIGC Detection Analysis
- GAHR’s text-side and image-side AIGC evaluations use separate unimodal tasks and remove the opposite modality and its dependent components.The Weibo26 construction pairs generated text with native images or generated images with native text.Source: §5.6 · p. 14
- GAHR achieves the best overall performance in both modality-specific generativity-detection settings.The text-side and image-side variants are GAHR-T and GAHR-I, respectively.Source: §5.6 · p. 14, Table 4 · p. 15, Table 5 · p. 15
- Accuracy improves across all six sample types when generativity labels are used, with more pronounced gains for fake-news types NFN, ICFN, and TCFN.The comparison is between veracity detection with and without generativity labels.Source: §5.6 · p. 15, Figure 4 · p. 15
- After GAHR processing, the six sample types appear more separated and real and fake samples overlap less in the t-SNE visualization.The original feature space shows greater entanglement and substantial real–fake overlap.Source: §5.6 · p. 15, Figure 5 · p. 16
5.7. Parameter Analysis
The parameter analysis finds that intermediate attention temperatures and a moderate generativity-loss weight perform better than extreme settings. The case study shows highlighted evidence associated with event content, while two errors illustrate difficulty distinguishing false narratives from surface coherence.
5.7. Parameter Analysis
- Generativity-supervision performance peaks at a moderate loss weight, while very small or large weights reduce performance; intermediate attention temperatures yield better Accuracy and fake-news and real-news F1-scores.The analysis covers AMG and Weibo21 for attention temperature and Weibo26 for the generativity-loss weight.Source: §5.7 · p. 16
5.8. Case Study
- In correctly classified cases, highlighted evidence concerns news subjects, entities, locations, event descriptions, and displayed objects.For real news, highlights mainly relate to the reported event and participants; for fake news, they focus on narrative-related textual or visual elements.Source: §5.8 · p. 16
- Two misclassified cases show that visually coherent image–text pairs grounded in false narratives can escape recognition.The ICRN case is factually consistent generated text; the TCFN case is classified as real despite a false narrative.Source: §5.8 · pp. 16–17
6. Conclusion
The conclusion presents Weibo26 as a dataset for multimodal fake-news detection when generative content participates in dissemination, with both veracity and generativity labels. It also reports that GAHR combines global judgment, local correction, and generativity detection, achieving competitive veracity results and strong generative-content detection results on the evaluated data.
6. Conclusion
- Weibo26 contains six sample types and provides both veracity and generativity labels for evaluating news involving generative content.The dataset is based on social-media news image–text pairs.Source: §6 · p. 17
- GAHR models the overall text–image relationship, uses local evidence to correct global judgment, and learns modality-specific generativity judgments on annotated data.The framework incorporates global judgment, local correction, and generativity detection in a unified model.Source: §6 · pp. 17–18
- GAHR achieves competitive veracity-detection performance on evaluated datasets and strong results on Weibo26 generative-content detection tasks.Source: §6 · p. 18